AI supplier document intake: extraction is not permission to import

A supplier document can contain useful item descriptions, prices and units, but extracting those fields is not the same as approving them for a construction cost library. A governed intake flow separates interpretation, review and execution so one uncertain line does not silently change future estimates.

Treat the file as evidence, not instruction

Keep the original document, supplier and receipt date. The file may contain notes or embedded text that should never be treated as system commands. AI can propose structured fields, but the application decides which schema and actions exist.

Mark unreadable pages and unsupported formats early. An incomplete extraction should remain incomplete rather than filling gaps with plausible values.

Show uncertainty line by line

Present the source text beside proposed article, unit, price, tax treatment and effective date. Highlight missing references, ambiguous units and conflicting totals. A reviewer needs to see why a value was proposed, not just a confidence colour.

Allow rejection or correction at line level. One clean item should not force acceptance of the whole document.

Enforce the review boundary

Enfin's current Alfie registry separates review_supplier_document_import from import_supplier_document. These supplier-document actions are exposed as web-only workflows. That boundary matters: a convenient chat or WhatsApp instruction must not be described as a supported route for this import.

The review prepares a decision. It does not write the supplier data by itself.

Apply permissions to the final action

The current actions require inventory-write permission. Check the authenticated organisation and user at execution time, not only when the file was uploaded. A person who can view the review may not be authorised to change the shared library.

Show exactly how many items will be created or updated before confirmation. Preserve the reviewer and action result in the audit trail.

Test downstream effects

After import, sample the changed items in a calculation. Verify units, purchase prices and margins. If a correction is needed, make it through a traceable update rather than editing the audit record.

Measure review quality, not automation volume

Track rejected lines, corrected units and imports reversed after use. These signals show whether extraction and review are working. A high count of processed documents says little if reviewers approve uncertain data under time pressure. Tighten the source, schema or confirmation step where recurring mistakes appear.

Meet Alfie as Enfin's governed construction execution layer, and read how to maintain a construction supplier price library.